Non-uniform Quantization for Neural Network Read Voltage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-volatile memory devices face challenges in maintaining data integrity due to noise introduced by program disturb and inter-cell interference, which affects the voltage levels over time, requiring improved methods to determine optimal read voltage thresholds.
Innovation Solution
Implementing deep neural networks with non-uniform quantization for power-of-two computations to reduce latency and inference time, allowing for the generation of updated read voltage values in memory devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If uniform quantization is used for power-of-two computations in DNN, then implementation is simpler, but latency and inference time increase
Solution Approach 1:
The patent applies non-uniform quantization specifically to the weight parameters of the DNN, while keeping the computational structure simple. This localized application of complexity to the right component (weights) achieves fast power-of-two computations without making the entire system complex
Solution Approach 2:
The patent changes the quantization parameter distribution from uniform to non-uniform, where quantization levels are adapted to the actual weight distribution. This parameter change enables more efficient representation that speeds up computation while maintaining accuracy
2Loss of time
If non-uniform quantization is used for power-of-two computations in DNN, then latency and inference time are reduced, but quantization complexity increases
Solution Approach 1:
The complex non-uniform quantization is applied only to the weight parameters, not to the entire computational process. The computational operations themselves remain simple power-of-two operations, localizing the complexity to where it provides maximum benefit
Solution Approach 2:
The non-uniform quantization levels are pre-computed and stored in lookup tables before inference. This preliminary preparation eliminates the need for complex real-time calculations during inference, achieving fast execution without runtime complexity
3Measurement precision
If standard DNN computation methods are used, then accuracy is maintained, but memory footprint and computational power requirements are high
Solution Approach 1:
The patent changes the parameter representation from standard floating-point to non-uniform quantized integers that represent power-of-two values. This parameter transformation reduces memory footprint and computational requirements while preserving accuracy through adaptive quantization levels
Solution Approach 2:
The patent replaces complex floating-point multiplication operations with simpler integer addition and bit-shifting operations. This substitution of computational mechanics reduces the computational power required while maintaining accuracy through the non-uniform quantization scheme
Data Source
AI summary
Devices, systems, and methods for improving operation of a memory device that uses a deep neural network (DNN), based on using non-uniform quantization for flexible power-of-two computations, are described. An example method includes receiving a plurality of initial weights of the DNN configured to determine a value of a read voltage associated with memory device, wherein the plurality of initial weights comprises at least one non-power-of-two quantized value or at least one floating point value. The method then aggregates the plurality of initial weights to generate a plurality of quantization functions, determines each of a plurality of quantized weights for a corresponding one of the plurality of quantization functions such that each of the plurality of quantized weights is a sum of powers-of-two, and configures the DNN to use the plurality of quantized weights to generate an updated value of the read voltage for retrieving information from the memory device.